1. StalemateBreaker: A Proactive Content-Introducing Approach to Automatic Human-Computer Conversation
Xiang Li,1 Lili Mou,1 Rui Yan,2 Ming Zhang1
1School of EECS, Peking University, China
2Natural Language Processing Department, Baidu Inc., China

2. Human-computer Conversation
One of the most challenging problems in artificial intelligence
The computer either searches or synthesizes a reply given an utterance (called query) issued by a user
Industrial products like Siri of Apple, Xiaobing of Microsoft, and Dumi of Baidu, etc.

3. Passive or Proactive Passive: To “respond” only
Proactive: To introduce new content when stalemate occurs
Human-human conversation statistics:

4. Mixed-initiative Systems in Vertical VS Open Domains
Vertical domain
Train95, AutoTutor, etc.
Feasible to manually design rules and templates
The content to be introduced is nearly certain
Open-domain
Users are free to say anything
Impossible to specify rules/templates in advance

5. Contributions
The first to address the problem of content introducing in open-domain conversation systems
A complete pipeline, involving when, what, and how to introduce new content
The Bi-PageRank-HITS reranking algorithm, emphasizing rich interaction between conversation context and candidate replies

6. Architecture Stalemate detection Named entity detection
Keyword filters like “Err”, “Errr”, etc.
Named entity detection
Named entities reflect users' interest
Candidate reply retrieval
Retrieve candidate replies by entities & conversation context
Selection by reranking
Candidate replies are reranked by a random walk-like algorithm

7. Process Flow
The system is built upon a conventional retrieval-based conversation system, which is typically passive

8. Reranking Algorithm
Importance of each query or reply, interaction between queries and replies
Bi-PageRank-HITS: combination of PageRank and HITS
Formulate queries and replies as a bipartite graph
Alternate between the following two steps
PageRank: Rank one side in the bipartite graph
HITS: Propagate information to the other side

9. Reranking Algorithm PageRank step: [·]: column normalization
M: similarity matrix of either queries or replies
x, y: prior distributions over queries and replies, uniformly initialized and updated after HITS
HITS step:
x: query scores, y: reply scores (see also PageRank)
Weight matrix is updated according to PageRank scores, given by
where φ(·,·) is a static, text-based relevance score

10. Reranking Algorithm
Global iteration over PageRank and HITS

11. Evaluation Dataset: sessions from real-world user conversation logs
Human evaluation: 1 point = good; 0 point = not good

12. Analysis
Parameter: textual information for queries does recommend important queries. On the contrary, textual information for replies is inimical.
Convergence: for the 10 randomly chosen samples, they typically converge quickly in 3--5 global iterations.

13. Case Study

14. Thanks for Listening
Q & A