The merits of translation at the word level or the concept level have long been a cause for debate among linguists. Some linguists sugges t that the two languages of a bilingual speaker share a common semantic system (Illes and Francis 1999; Ikeda 1998) and hence translation is carried out at the concept level.

Meanwhile, there has been a gradual convergence of statistical and transfer approaches in machine translation recently (Wu 2003). Statistical MT systems are based on a stochastic mapping between lexical item s, assuming the underlying semantic transfer is hidden. Transfer systems use explicit lexical, syntactic and semantic transfer rules. Consequently, cognitive scientists and computational linguists alike have been interested in the study of semantic mapping between languages (Ploux and Ji, 2003, Dorr et al., 2002, Ngai et al., 2002, Boas 2002, Palmer and Wu, 1995). We propose to au tomatically construct a bilingual lexical semantic network with word sense and semantic role mapping between English and Chinese, simulating the “concept lexicon”, suggested b y cognitive scientists, of a bilingual person. Figure 1. BiFrameNet lexicon and example

sentence induction

The linguists-defined ontologies–-FrameNet (Baker et al., 1998), HowNet (Dong and Dong, 2000), and bilingual dictionaries are the basis for the induction of the mapping. We automatically estimate the semantic transfer likelihoods bet ween English FrameNet lexical entries and the Chinese word senses in HowNet, and align those frames and lexical pairs with high likelihood values. In addition, we propose to induce Ch inese example sentences automatically to match English annotated sentences provided in the FrameNet. The BiFrameNet thus induced provides an additional resource for machine-aided or machine translation systems. It can also serve as a reference to be compared to cognitive studies of the translation process.

Ploux and Ji, (2003) proposed a spatial model for matching semantic values between French and English. Palmer and Wu (1995) studied the mapping of change-of-state English verbs to Chinese. Dorr et al. (2002) described a technique for the construction of a Chinese-English verb lexicon based on HowNet and the English LCS Verb Database (LVD). They created links between HowNet concep ts and LVD verb classes using both statistics and a manually constructed “seed mapping” of thematic classes between HowNet and LVD. Ngai et al. (2002) employed a word-vector based approach to create the alignment between WordNet and HowNet cl asses without any manual annotation. Boas (2002) outlined a number of issues surroundin g the planning and design of GermanFrameNet (GFN), a bilingual FrameNet dictionary which, when complete, will have a corpus-based German lexicon following the FrameNet structure.

This paper is organized as follows: Section 2 describes the algorithm for estimating transfer relations between FrameNet and HowNet structures. Section 3 presents our method for selecting BiFrameNet example sentences for a particular frame and automatically inducing semantic role annotations. We conclude in Section 4, followed by a discussion in Section 5.

Dorr et al. (2002) uses a manual seed mapping of semantic roles between FrameNet and LVD to induce a bilingual verb lexicon. In this paper, we propose a method of automatically mapping the English FrameNet lexical entries to HowNet concepts, resulting in the BiFrameNet ontology. We also make use of two bilingual English-Chinese lexicons for this induction. In this section 2, we use an example FrameNet lexical entry “beat.v” in the “cause_harm” frame to illustrate the main steps of our algorithm. µ : lexical entry in FrameNet; ν : concept in HowNet F : FrameNet frame; H :HowNet category L : links from FrameNet to HowNet L F : links of the frame F; T µ : translations of µ R is a ranked list; R [ k ] means the top- k element V F :possible ν linked to µ∈ F Λ F :HowNet categories related to frame F δ ( x ) : binary function, return 1 if input is true;

otherw ise return 0. For each µ T1 µ ={translation s of µ in HowNet} T2 µ ={translations of µ in dictionary} T = T 1 U T 2

The Berkeley FrameNet database consists of frame-semantic descriptions of more than 7000 English lexical items, together with example sentences annotated with semantic roles (Baker et al., 1998). There is currently no frame semantic representation of Chinese. However, the Chinese HowNet (Dong and Dong 2000) represents a hierarchical v iew of lexical semantics in Chinese.

FrameNet is a collection of lexical entries grouped by frame semantics. Each lexical entry represents an individual word sense, and is associated with semantic roles and some annotated sentences. Lexical entries with the same semantic roles are grouped into a “frame” and the semantic roles are called “frame elements”. For example:

HowNet is a Chinese ontology with a graph structure of word senses called “concepts”, and each concept contains 7 fields including lexical entries in µ µ µ L ← L U {( µ , ν ) | ν . W_C ∈ T µ , ν . G_C = ν . POS } For each F For each H V F ={ ν |(u, ν ) ∈ L , µ ∈ F } f ( H ) = ∑ δ ( ν ∈ V F , v ∈ H ) ν R is the ranked list of H sorted by f ( H ) Λ F = R [1] U R [2]... U R [ N ] Λ F ← Λ F U { R [ k ] | Sim ( R [ l ], R [ k ]) > threshold ,

l = 1,.. N , k = N + 1,..} Figure 2. BiFrameNet ontology induction L F = {( µ , ν ) | µ ∈ F , ν ∈ V F I Λ F } Chinese, English gloss, POS tags for the word in Chinese and English, and a definition of the concept including its category and semantic relations (Dong and Dong, 2000). For example, one translation for “beat.v” is 打:

Whereas HowNet concepts correspond roughl y to FrameNet lexical entries, its semantic relations do not correspond directly to FrameNet semantic roles.

We use the bilingual lexicon from HowNet and LDC dictionary to first create all possible m appings between FrameNet lexical entries and HowNet concepts whose part-of-speech (POS) tags are the same. Here we assume that syntactic classification for the majority of FrameNet lexical entries (i.e. verbs and adjectives) are s emantically motivated and are mostly preserved across different languages. For example “beat” can be translated into { 搥, 败, 冲击, 出手, 难倒, 骗取, 赢, 战败… } in HowNet and { 打, 打败, 捣, 敲打, 赢… } in the LDC English-Chinese dictionary. “beat.v” is then linked to all HowNet concepts whose Chinese word/phrase is one of the translations and the part of speech is verb “v”.

At this stage, each FrameNet lexical entry has links to multiple HowNet concepts and categories. For example, “beat.v” in “cause_harm” frame is linked to “ 打 ” in both the “beat” category and the “associate” category (as in“ 打电话 /make a phone call”). We need to choose the correct HowNet concept (word sense). Many word sense disambiguation algorithms use contextual words in a sentence as di sambiguating features.

In this work, we make use of contextual lexical entries from the same semantic frame, as illustrated above. In this example, the “cause_harm” frame contains two lexical entrie s—“beat.v” and “strike.v”. From the previous step, “beat.v” and “strike.v” is each linked to a number of Chinese candidates. “beat.v” is linked to “ 打” with membership in two different HowNet categories, nam ely “ 打 |beat” and “ 交往 |associate” . To disambiguate between the above these 2 candidate categories, we make use of the other lexical entries in “ cause_harm”, in this case “strike.v” which is linked to “ 捶 ”, in the “ 打 | beat” HowNet category. Now, “ |beat” receives two votes (from “ ” and from “ ”), and “ |associate” only one (from “ ”). We therefore choose the HowNet category “ |beat” to be aligned to the frame “cause_harm”, and eliminate the sense of “ 打 ”in the “ 交往 | associate” category. Consequently, “beat.v” in “cause_harm” is linked to all HowNet concepts that are translations of “beat” which are verbs, and which also belong to the How Net category “ |beat” (but not “ |associate”).

In our example, HowNet concepts under two HowNet categories—“beat” and “damage” are linked to the “cause_harm” frame in FrameNet. Only the concepts in the top N categories are c onsidered as correctly linked to the lexical entries in the “cause_harm” frame. We heuristically chose N to be three in our algorithm.

Using frame context alone in the above step can effectively prune out incorrect links, but it also prunes some correct links whose HowN et categories are not in the top three categories. In this next step, we aim to recover this kind of pruned links by finding other categories with high similarity to the chosen categories. We introduce the category similarity score (Liu and Li, 2002), which is based on the HowNet taxonomy distance:

Sim(category1,category2) = α +d

Where d is the path length from category1 to category2 in the taxonomy. α is an adjusting parameter, which controls the curvature of th e similarity score. We set α = 1.6 in our work following the experiment results in Liu and Li (2002) . If the similarity of category p and one of the top three categories is higher than a threshold t, the category p is also considered as a vali d category for the frame. α

In our example, some valid categories, such as “firing| 射击 ” is not selected in the previous step even though it is related to the “cause_harm” frame. Based on the HowNet taxonomy, the similarity score between “firing| ” and “beat| ” is 1.0, which is above the threshold set. Hence, “firing| ” is also chosen as a valid category and the concepts in this category are linked to the “beat.v” lexical entry in the “cause_harm” frame. However, using taxonomy distance can cause errors such as 打 in the “weave” category to be aligned to “beat.v” in the “cause_harm” frame.

We evaluate our work by comparing the results to a manually set golden standard of transfer links for some lexical entries in FrameNet, and use the precision and recall rate as evaluation criteria. Manual evaluation of all lexical entries is a slow process and is currently still on-going. However, to show the lower bound of the system performance, we chose FrameNet lexical entries with the highest number of transfer links to HowNet concepts as the test set. Since each link is a word sense, these lexical entries have most ambiguous translations. Since the number of lexical entries in a FrameNet parent frame (i.e. frame size) is an important factor in the disambiguation step, we analyze our results by distinguishing between “small frames” (a frame with less than 5 lexical entries) and “large frames”. 24% of the frames are “small frames”. Referring to Tables 2 and 3, we can see a weighted average of (0.649*0.24+0.874*0.76) =82% F-measure. Table1. Lexical entries test set Table 2.Performance on large frames Table 3. Performance on small frames Table 4. Average performance on “beat.v” at

each step of the algorithm

Table 4 shows the system performance in each step of the alignment between the FrameNet “beat.v” to HowNet concepts with the final F-measure at 89.72.

example

In the second stage of our proposed work, we aim to automatically induce Chinese example sentences that are appropriate for each semantic frame. Together with English example sentences that already exist in the English FrameNet, they form part of the BiFrameNet, and serve to provide concrete examples of bilingual usage of semantic roles. They can be used either as a resource for machine-aided translation or training data for machine translation.

FrameNet is a collection of over 100-million words of samples of written and spoken language from a wide range of sources, including British and American English. All the example sentences are chosen by linguists for their representative-ness of particular semantic roles, grammatical functions, and phrase type. The current FrameNet contains on average 30 annotated example sentences per predicate, which is still inadequate for automatic semantic parsing systems (Fleischman et al., 2003). Each FrameNet example sentence contains a predicate. The semantic roles of the related frame elements are manually labeled. The syntactic phrase type (e.g. NP, PP) and their grammatical function (e.g. external argument, object argument) are also labeled. An example annotated sentence containing the predicate “beat.v”, in the “ cause_harm” frame, is shown below:

In order to provide a representative set of Chinese example sentences automatically for a particular frame, our method must fulfill the following criteria: 1) It must find real sentences occurring naturally in

Chinese texts;

It should find sentences that cover as many

different usage and domain as possible;

It must find sentences that have the same

semantic roles as the English example sentences; 2) 3) Figure 3. BiFrameNet example sentence induction 4) It should require no manual annotation of any

kind.

There are at least three different (semi-)automatic approaches for m ining Chinese example sentences:

i) Translate all English e xample sentences into Chinese by automatic means, and annotate t he semantic roles by word alignment;

This approach is not appropriate because machine translation can be erroneous and this method does not satisfy criteria (1) and (2).

ii) Construct an English semantic parser and a Chinese parser independently, and use them to annotate the sentences in a sentence aligned, parallel corpus;

Apart from the high cost of building two semantic parsers, which itself requires semantically annotated Chinese data; it would be necess ary to create artificial links between independ ent human annotations manually .

iii) Mine Chinese sentences from a monolingual corpus that are syntactical ly similar to the English example sentence, and in duce semantic roles from the syntactic transfer function between English an d Chinese.

This is the approach we take. Inspired by previous work on syntax-driven semantic parsin g (Gildea and Jurafsky, 2002; Fleischman et al., 2003), and syntax-based machine translation (Wu, 1997; Cuerzan and Yarowsky , 2002), we postulate that syntactically similar sentences with the same predicate also share similar semantic roles. In this paper, we present our first experiments on i nducing semantic roles based on shallow syntactic information. We mine Chinese example sentences from naturally occurring monolingual corpus, and rank them by their syntactic similarity to our English example sentences. A dynam ic programming algorithm then annotates the aligned syntactic units with the same semantic ro les. The example Chinese sentences are not translations of the English sentences. Therefore, the set of example sentences within a frame is enriched , providing better coverage for MT and CLIR sy stems.

Since frames are disjoint, we propose a method for finding example sentences one frame at a time. In this paper, we focus on finding Chinese example sentences for the largest frame “cause_harm” and the main semantic roles in this frame—“agent”, “predicate” and “victim” 1 .

For each English lexical entry and its target translation candidates in the BiFrameNet, we first extract sentences that contain the translation candidates from a large Chinese monolingual corpus. Figure 4 shows some initial Chinese example sentence candidates under “beat.v”. There are many sentences that do not have the “agent-predicate-victim” structure. Our next step is to find the Chinese sentences that have the “agent”, ”predicate” and ”victim” semantic roles and annotate them automatically. Figure 4. Some Chinese example sentence and glosses

Among all the Chinese sentences containing t he target predica te words, we need to identify those that contain the same semantic roles as those of the English example sentences in FrameNet. Current automatic semantic parsing algorit hms (Gildea and Jurafsky 2003, Fleischman et al., 2003) are all based on syntactic parse trees showing a close coupling of semantic and syntactic structures.

Without carry ing out full syntactic parsing of the Chinese sentences, we postulate that the semantic roles of a sentence are generated by the underlying shallow syntactic structure of the sentence such as POS tag sequences. We therefore focus on finding bilingual sentence pairs that are comparable in POS structure, though not n ecessarily having any lexical comparability. Note that this constitutes onl y a subset of all possible Chinese example sentences for each frame. The expansion of this set remains the objective of our future research

Given an English example sentence, its semantic role sequence, and its POS tag sequence; and a set of Chinese sentences and their POS tag sequence, we use a dynamic programming m ethod (Figure 5) to find the Chinese sentenc e whose POS sequence is most likely to be generated from the English POS sequence, and the alignment path. The Chinese word aligned to the English word will assu me the latter’s semantic role. Figure 5. Dynamic programming (DP) alignment We train σ ( ec , ) in Figure 5 from a sentence aligned, POS tagged, parallel corpus (Hong Kong News), and a bilingual dictionary. For each bilingual word pair in the dictionary, we estimate the prior distributions of the POS tags of the Chinese words from the Chinese side of the parallel corpus, and that of the English words from the English side. A V x W POS tag “confusion matrix” is generated, where V is the vocabulary of the Chinese POS tags, and W is the vocabulary of the English POS tags. Table 5 shows some example English-Chinese POS mapping and Figure 6 shows some example annotated sentences in Chinese. Table 5. Example POS tag transfer Figure 6. Example Chinese annotated sentences

We estimate the syntactic POS transfer probabilities from the HK News Corpus. We use two state-of-the-art POS taggers—a maximum entropy based English POS tagger (Ratnaparkhi, 1996), and an HMM based Chinese POS tagger. 2 We perform two sets of experiments: (1) For each example English sentence in the “cause_harm” frame from FrameNet, we extract a corresponding Chinese sentence annotated with t he same semantic roles; (2) rank all the Chinese sentences that have been aligned to the English sentences by alignment score . The highest ranking Chinese sentenc es are used for the BiFrameNet. Table 6 shows that the average annotation accuracy of all top Chinese sentence candidates for each English example sentence is 68%. Table 7 shows that t he annotation accuracy of the top 100 Chinese example sentences, sorted by DP score, is 71.8%. Table 6. Annotation accuracy of the selected

Chinese sentences

Table 7. Annotation accuracy of the top 100 Chinese sentences with the highest DP alignm ent

scores

We have presented a first quantitative and automatic approach of constructing a bilingual lexical semantic resource—the BiFrameNet. BiFrameNet consists of mappings between FrameNet semantic frames and HowNet concepts, as well as English and Chinese example sentences for a particular frame, with annotated semantic roles in the English FrameNet labels. Evaluation results show that we achieve a promising 82% average F-measure on lexical entry alignment, for the most ambiguous lexical entries; and a 68-72% accuracy in Chinese example sentence induction, for the largest frame. The initial results are available at http://www.cs.ust.hk/~hltc/BiFrameNet and will be updated as further improvements and evaluations are implemented.

There are a number of possible directions for future work. One obvious extension is to use syntactic parse tree representations instead of POS sequences in example sentence alignment. Second, there are many other Chinese sentences that share the same semantic roles, but not the same POS sequences, which are not included. Using additional features to correctly identify these sentences and the constituent semantic roles is a topic of our ongoing research. Moreover, we note that Chinese is a highly idiomatic and metaphoric language. Compounded by the ambiguity of word boundaries, many predicate usages in Chinese are highly unexpected. It is worth considering using other Chinese linguistic resources to enhance the example sentence extraction and annotation. Finally, BiFrameNet needs to be further evaluated and manual post-processing is perhaps required.

We expect the final complete BiFrameNet, in addition to the various FrameNet and PropBank resources being developed manually, will be a valuable resource for statistical and interlingua transfer-based MT systems, as well as to human translators in an machine-aided translation scenario. We are also motivated to investigate the relationship between our results and those of semantic mapping models proposed by cognitive scientists.

This work is partly supported by grants CERG# HKUST6206/03E and CERG#HKUST6213/02E of the Hong Kong Research Grants Council. References Collin F. Baker, Charles J. Fillmore and John B. Lowe.

(1998).The Berkeley FrameNet project. In

Proceedings of the COLING-ACL , Montreal,

Canada. Hans C. Boas. Bilingual FrameNet Dictionaries for

Machine Translation. In Proceedings of the Third

International Conference on Language Resources

and Evaluation. Las Palmas, Spain. Vol. IV:

1364-1371 2002. Silviu Cucerzan and David Yarowsky. Bootstrapping a

multilingual part-of-speech tagger in one person-day.

In Proceedings of the Sixth Conference on Natural

Language Learning (CoNLL). Taipei, Taiwan. 2002. Dong, Zhendong., and Dong, Qiang. HowNet [online

2002]. Available at

http://www.keenage.com/zhiwang/e_zhiwang.html Bonnie J. Dorr, Gina-Anne Levow, and Dekang

Lin.(2002).Construction of a Chinese-English Verb

Lexicon for Machine Translation. In Machine

Translation, Special Issue on Embedded MT , 17:1-2. Michael Fleischman, Namhee Kwon and Eduard Hovy.

Maximum Entropy Models for FrameNet

Classification. In Proceedings of ACL 2003 ,

Sapporo. Daniel Gildea and Daniel Jurafsky.(2002).Automatic

Labeling of Semantic Roles. In Computational

Linguistics, Vol 28.3: 245-288. Judy Illes and Wendy S. Francis. Convergent cortical

representation of semantic processing in bilinguals.

In Brain and Language , 70(3):347-363, 1999. Liu Qun and Li, Sujian. Word Similarity Computing

Based on How-net. In Computational Linguistics

and Chinese Language Processing ， Vol.7, No.2,

August 2002, pp.59-76 Grace Ngai, Marine Carpuat, and Pascale Fung.

Identifying Concepts Across Languages: A First

Step towards a Corpus-based Approach to

Automatic Ontology Alignment. In Proceedings of

COLING-02 , Taipei, Taiwan. Martha Palmer and Wu Zhibiao . Verb Semantics for

English-Chinese Translation . In Machine

Translation 10: 59-92, 1995. Sabine Ploux and Hyungsuk Ji. A Model for Matching

Semantic Maps between Languages (French/English,

English/French). In Computational Linguistics

29(2):155-178, 2003. Adwait Ratnaparkhi. A Maximum Entropy

Part-Of-Speech Tagger. In Proceedings of EMNLP

2003, May 17-18, 1996. University of Pennsylvania Satoko Ikeda. Manual response set in a stroop-like task

involving categorization of English and Japanese

words indicates a common semantic representation.

In Perceptual and Motor Skills , 87(2):467-474,

1998. Dekai Wu. Stochastic inversion transduction grammars

and bilingual parsing of parallel corpora . In

Computational Linguistics 23(3):377-404, Sep 1997 Dekai Wu. The HKUST leading question translation

system. MT-Summit 2003. New Orleans, Sep 2003.