Generally speaking, sentences are constituted by substantives and functional words. Domain specific terms (terms for short) are more likely to be domain substantives. Words immediate before and after these terms, called predecessors and successors of the terms, are likely to be either functional words or other general substantives connecting terms. These predecessors and successors can be considered as markers of terms, and are referred to as term delimiters in this paper. In contrast to terms, delimiters are relatively stable and domain independent. Thus, they can be extracted more easily. Instead of looking for features associated with terms as in other works, this paper looks for features associated with term delimiters. That is, term delimiters are identified first. Words between delimiters are then taken as term candidates.

The proposed delimiter identification based algorithm, referred to as TCE_DI (Term Candidate Extraction - Delimiter Identification), extracts term candidates from a domain corpus by using a delimiter list, referred to as the DList. Given a DList, the algorithm TCE_DI itself is straight forward. For a given character string CS (CS = C1C2...Cn) shown in Figure 1, where Ci is a Chinese character. Suppose there are two delimiters D1 = Ci1...Cil and D2 = CJ1^CJm in CS where D1 e DList and D2 e DList. The string CS is then segmented to five substrings: C1.Cib, Ci1...Cil and CJ1 — CJm are delimiters, C1 — Cib, Cia...Cjb, and CJa...Cn are regarded as term candidates as labeled by TC1, TC2 and TC3 in Figure 1, respectively. If there is no delimiter contained in CS, the whole string C1C2.Cn is regarded as one term candidate.

DList can be obtained either from a delimiter training corpus or from a given stop word list. Given a delimiter training corpus, CorpusTraining, normally a domain specific corpus, and a domain lexicon Lexicon, DList can be obtained based on the following algorithm, referred to as DListExt (DelimiterList Extraction Algorithm).

Step 1 : For each term Ti in Lexicon, mark Ti in

CorpusTraining as a non-divisible lexical unit. Step 2: Segment remaining text in CorpusTraining. Step 3: Extracts predecessors and successors of

all Ti as delimiter candidates. Step 4: Remove delimiter candidates that are

contained in a Ti in Lexicon. Step 5: Rank delimiter candidates by frequency and the top NDI number of items are considered delimiters.

The DList Ext algorithm basically use known terms in a domain specific Lexicon to find the delimiters. It can be shown in the experiments later that Lexicon does not need to be comprehensive. Even if a small training corpus, CorpusTraining, is not available in a language without sufficient domain specific NLP resources, a stop-word list produced by experts or from a general corpus can serve as DList directly without using the DList Ext algorithm.

In a domain corpus, some sentences are domain relevant sentences which contain more domain specific information whereas others are general sentences which contain less domain information. A domain specific term is more likely to be contained in domain relevant sentences, which means that domain relevant sentences and domain specific terms have a mutually reinforcing relationship. A novel algorithm, referred to as TVLinkA (Term Verification -Link Analysis) based the Hyperlink-Induced Topic Search (HITS) algorithm (Kleinberg, 1997) originally proposed for information retrieval, is proposed using link analysis to calculate the relevance between term candidates and the sentences in domain specific corpora for term verification.

In TV LinkA, a node p can either be a sentence or a term candidate. If a term candidate TermC is contained in a sentence Sen of the corpus CorpusExtract where the candidates were extracted, there is a directional link from Sen to TermC. This way, a graph for the candidates and the sentences in CorpusExtract can be constructed and the links between them indicate their relationships. A good hub in CorpusExtract is a sentence that contains many good authorities; a good authority is a term candidate that is contained in many good hubs. Each node p is associated with a non-negative authority weight w(p)A and a non-negative hub weight w(p)H . Link analysis in TV LinkA makes use of the relationship between hubs and authorities via an iterative process to maintain and update authority/hub weights for each node of the graph.

Let VA denote the authority vector (w(p1)A, w(p2)A,..., w(pn)A) and V denote the hub vector (w(p1)H, w(p2)H,..., w(pn)H), where n is the sum of the total number of sentences and the total number of term candidates. Given weights VA and V with a directional link p —q, the I operation(an in-pointer to a node) and the O operation(an out-pointer to a node) update w(q)A and w(p)H as follows.

Let k be the iteration termination parameter and z be the vector (1, 1, 1,., 1) , and VA and VH are initialized to V0A = V0H = z. Hubs and authorities can then be calculated as follows.

obtaining new V,A.

Apply the O operation to (V'A , VHX ), obtaining new V,H .

Normalize V]A by dividing the normalization factor (w' (p)A ) to obtain ViA .

Normalize V]H by dividing the normalization factor (w' (p)H ) to obtain ViH .

Return (VkA , VkH)

In CorpusExtract, term candidates with high authority in a few documents are likely to be domain specific terms whereas candidates with high authority in many documents are more likely to be commonly used general words. Based on this observation, the termhood of each candidate term TermC, denoted as TermhoodC, is calculated according to formula (3) defined below.

where w(C)jA is the authority of TermC in a document Dj of CorpusExtract, |D| is the total number of documents in CorpusExtract and DFC is the total number of documents in which TermC occurs. TermCs are then ranked according to their termhood values TermhoodC, and the top ranked NTCList candidates are considered terms. NTCList is an algorithm parameter to be determined experimentally.

For comparison, a statistical based term candidate extraction algorithm, TCE SEF&CV with the best performance in (Ji and Lu, 2007) using both internal association and external strength, is used as the reference algorithm for the evaluation of TCE_DI. A statistics based term verification algorithm, TV_ConSem (Ji and Lu, 2007) using semantic information within a context window is used for the evaluation of TV_LinkA. LexiconPKU is also used in TV_ConSem. Two popular methods integrated without division of candidate extraction and verification steps are used for comparison. The first one is based on TF-IDF (Salton and McGill, 1983; Frank et al., 1999). The second one is a supervised learning approach based on a SVM classifier, SVMlight (Jo achims, 1999). The features used by SVMiight are shown in Table 3. Two training sets are constructed for the SVM classifier. The first one includes 3,337 positive examples (LexiconIT) and 5,950 negative examples extracted from CorpusIT_Smaii. The second one includes 394 positive examples (LexiconLegai) and 28,051 negative examples extracted from CorpusLegai_smaii.

No.

_Feature Explanation_

Percentage of the Chinese characters occurred in LexiconDomainFrequency in the domain corpus Frequency in the general corpus Part of speech

The length of Chinese characters in the candidate

The length of non-Chinese characters in the candidate Contextual evidence

Figure 2 shows the performance of the proposed TCE_DI and TVLinkA for term extraction compared to the reference algorithms for IT domain using CorpusIT_Large. TCE_DIIT and TCE_DIiegai indicate TCE_DI using extracted delimiter lists DListIT and DListLegal with NDI = 500, respectively. TCE_DISW simply uses the stop word list DListSW.

As shown in Figure 2, term extraction based on TCE_DIIT combined with TV LinkA gives the best performance. It achieves 75.4% precision when the number of extracted terms NTCList reaches 5,000. The performance is 9.6% and 29.4% higher in precision compared to TF-IDF and TCESEF&CV combined with TV ConSem, respectively. These translate to improvements of precision of over 14.8% and 63.9%, respectively.

When applying the same TV LinkA algorithm for term verification, TCE_DI using different delimiter lists provide 24% better performance on average compared to the TCESEF&CV algorithm which translates to improvement of over 47%. The result from using delimiters of legal domain (DListLegai) to data in IT domain (as shown in TCE_DIiegai) is better on average than using a simple general stop word list. It should be noted, however, that TCE_DISW still performs much better than the reference algorithms, which means that delimiter based term candidate extraction algorithm can improve performance even without any domain specific training. When applying the same TCE_DIIT algorithm in term candidate extraction, TV_LinkA provides 10% higher performance compared to the TV_ConSem algorithm which translates to improvement of over 15.3%. It is important to point out that TV_LinkA using only the stop word list without any domain specific knowledge performs better than TV_ConSem using a large domain lexicon. In other words, delimiter based extraction with link analysis use much less resources and still improve performance of TV_ConSem.

The performance of TCE_DIIT or TCE_SEF&CV combined with TV_ConSem have an upward trend when more terms are extracted which seems to be against intuition. The principle of the TV_ConSem algorithm is that a candidate is considered a valid term if a majority of its context words already appear in the domain lexicon. General words are more likely to be ranked on top because they are commonly used which explains the low performance of TV_ConSem in the lower range of NTCList. When NTCList increases, more domain terms are included. Thus, there is an upward trend in precision. But, the upward trend reverts at around 4,500 because the measurement in percentage is too low to distinguish valid terms from non-term candidates.

It is also interesting to point out that the simple TF-IDF algorithm which was rarely used in Chinese term extraction performs as well as the SVM classifier. The main reason is that the test corpus consists of academic papers. So, many terms are consistent and repeated a lot of times in different documents which accords with the idea of TF-IDF. Thus, TF-IDF performs relatively well because of the high-quality domain corpus. However, TF-IDF, as a statistics based algorithm suffers from similar problem as others based on

Extracted Terms (NTCList)

statistics. Thus it does not perform as well as the proposed TCE_DI and TVLinkA algorithms.

Figure 3 shows that the proposed algorithms achieve similar performance on the legal domain. TCE_DILegal combined with TVLinkA perform the best. The result from using IT domain delimiters (DListIT) in legal domain as shown in TCE_DIIT is better on average than using the general purpose stop list. This further proves that extracted delimiter list even from a different domain can be more effective than a general stop word list. When applying the same TVLinkA algorithm for term verification, TCE_DI using different delimiter lists are better than all the reference algorithms. Without large lexicon in Chinese legal domain, the TVConSem algorithm does not even work. TV LinkA using no prior domain knowledge for term verification still achieves similar improvement compared to that of the IT domain where a comprehensive domain lexicon is available.

There are three main reasons for the performance improvements of the proposed TCE_DI and TV_LinkA algorithms. Firstly, the delimiters which are mainly functional words (e. g. "iS"(at/in), "gK"(or)) and general substantive (e.g. ";si|"(be), "(adopt)) can be extracted easily and are effective term boundary markers since they are quite domain independent and stable. Secondly, the granularity of domain specific terms extracted the proposed algorithm is much larger than words obtained by word segmentation. This keeps many noisy strings out of the term candidate set. Thus, the proposed delimiter based algorithm performs much better over segmentation based statistical methods. Thirdly, the proposed approach is not as sensitive to term frequency as other statistical based approaches because term candidates are identified without regards to the frequencies of the candidates. In the TV_LinkA algorithm, terms are verified by calculating the relevance between candidates and the sentences instead of the distributions of terms in different types of documents. Terms having low frequencies can be identified as long as they are in domain relevant sentences whereas in the previous approaches including TF-IDF, terms with less statistical significance are weeded out. For example, a long IT term "MM-t^it^fc" (Hierarchical storage system) with a low frequency of 6 is extracted using the proposed approach. It cannot be identified by TF-IDF since the statistical information is not significant. This term cannot be extracted by the segmentation based algorithms either because general segmentor split long terms into pieces making them difficult to be reunited using term extraction techniques.

It is interesting to know that the proposed approach not only achieves the best performance for both domains, it also achieves second best when using extracted delimiters from a different domain. The results confirm that delimiters are quite stable across domains and the relevance between candidates and sentences are efficient for distinguishing terms from non-terms in different domains. In fact, the proposed approach can be applied to different domains with minimal training or no training if resources are limited.

As LexiconPKU is the only ready-to-use domain lexicon, the evaluation on new term extraction is conducted on CorpusIT_Large only. Figure 4 shows the evaluation of the proposed algorithms compared to the reference algorithms in terms of RNTE, the ratio of new terms among all identified terms.

It can be seen that the proposed algorithms TCE_DIIT combined with TV_LinkA is basically the top performer throughout the range. It can identify 4% (with respect to TCESEF&CV +TVConSem) to 27% (with respect to TF-IDF) more new terms when NTCList reaches 5,000 which translate to improvements of over 9% to 170%, respectively. The second best performer is TCE_DIlegal combined with TV_LinkA using delimiters of legal domain. In fact, it only underperforms in the lower range of NTCList compared to TCE_DIIT. When NTCList reaches 5,000, their performance is basically the same. However, the TCE_DISW algorithm using stop words performs much worse than using extracted delimiter lists as shown for TCE_DIIT and TCE_DIlegal. In the TCE_DI algorithm, character strings are split by delimiters and the remained parts are taken as term candidates. Generally speaking, if a new term contains a delimiter or a stop word as its component, it cannot be identified correctly. Consequently, if a new term contains a stop word as its component, it cannot be extracted correctly using TCE_DISW. However, new terms are less likely to contain delimiters because the delimiter extraction algorithm DList_Ext would not consider a component as a delimiter if it is contained in a term in LexiconDomain. Consequently, TCE_DISWis less adaptive to domain specific data compared to TCE_DIIT and TCE_DIlegal. That is also why TCE_DISW picks up new terms much more slowly.

Extracted Terms (N )

for New Term Extraction

It is interesting to know that TCE_DIIT combined with TV_ConSem identifies more new terms in the low range of NTCList. In the TV_ConSem algorithm, the major information used for term verification is the percentage of the context words appear in the domain lexicon. As discussed earlier in Section 3.2, TV_ConSem ranks commonly used general words higher than others which leads to the low precision of TV_ConSem for term extraction. A new term faces a similar scenario because more of its context words occur in the domain lexicon than that of other terms. Thus, new terms are actually ranked higher than other terms in TV_ConSem which explains its higher ability to identify new terms in the low range of NTCList. However, its performance drops in the high range of NTCList because the influence of context words diminishes in terms of percentage in the domain lexicon to distinguish terms from non-terms. Figure 4 also shows that TF-IDF and SVM perform the worst in new term extraction compared to other algorithms. TF-IDF has relatively low ability to identify new terms since new terms are not widely used and they do not repeat a lot of times in many documents. As SVM is sensitive to training data, it is naturally not adaptive to new terms.

All current Chinese term extraction algorithms rely on segmentation with comprehensive lexical knowledge and yet Chinese segmentation algorithms have the OOV (out of vocabulary) problem. This makes Chinese term extraction particularly vulnerable to new term extraction. The proposed approach, on the other hand, is based on delimiters which is more stable, domain independent, and OOV independent. Figure 4 shows that TCE_DI and TV_LinkA using minimal training from different domains can extract much more new terms than previous techniques. In fact, the proposed approach can serve as a much better tool to identify new domain terms and can be quite effective for domain lexicon expansion.